Two professors went from near zero research hours to between eight and twelve.

Two quite different people who share a calendar problem. The professor: nine SWS of Lehrdeputat, between four and fifteen doctoral candidates, a Prüfungsausschuss, a Berufungskommission and the Fakultätsrat, six to ten manuscripts reviewed a year unpaid, and two or three grant proposals because the chair runs on Drittmittel. They were hired for their research and they do research on Sundays, if at all. And the doctoral candidate: year three of a WissZeitVG contract that ends in year four, 900 unread PDFs, three written chapters that no longer agree with each other, and a pipeline that cannot reproduce its own figures.

For professors it is the proposal. A DFG Sachbeihilfe needs the state of the art, preliminary work, the work programme, budget justification, the CV, the publication list in the DFG’s own structure, a data management plan, an ethics statement and equal opportunity measures. Perhaps sixty per cent of that text already existed, in three previous proposals, in slightly wrong formats — and it gets rewritten anyway, over five evenings-and-weekends weeks, because finding the old version feels harder than retyping it. For candidates it is six hundred abstracts to screen, and later the worse question: which version of the script, on which version of the data, produced Figure 3?

What the machine takes

  • Screening literature: pulling everything new in a defined space and sorting by relevance to a stated question
  • Deduplicating a reference library, repairing metadata, resolving every citation to a DOI
  • Reformatting a publication list into DFG, ERC, NSF, Europass or a university’s house structure
  • Maintaining the reusable proposal core and reshaping it to a call’s structure and page limits
  • Tracking calls, Zwischenberichte, Abschlussberichte, review commitments and exam dates
  • Keeping the supervision record for each doctoral candidate
  • Screening appointment applications against the published formal criteria for completeness
  • Keeping the reproducibility spine: which data, which script version, which figure

What stays yours

  • The research question. The hypothesis. The interpretation
  • Whether a result means anything
  • The assessment of a person — in a reference, a viva, a Berufungsverfahren
  • What to review, and what a review concludes

What I build

The proposal core
One maintained store: the CV in every format ever demanded, the publication list, the group description, infrastructure, the preliminary work paragraphs per research line with their figures and citations, past project outcomes, the data management plan template, the equal opportunity measures. Every item dated and sourced. About two weeks to build, mostly extraction from three old proposals, and it is the last time those paragraphs are ever written from nothing.
The call shaper
A call’s structure, headings and page limits go in. Out comes a skeleton with the core material already placed at roughly the right length, and every section needing original thought marked empty and labelled. Five weeks becomes about ten days, and almost all ten are spent on the work programme — which is the part that gets funded.
The supervision record
One private file per doctoral candidate: what was agreed, what happened, milestones against the Betreuungsvereinbarung, obstacles raised. Ninety seconds of dictation after each meeting. The piece I would install first if I could install only one — because meetings currently begin with both people trying to remember, reference letters two years later are written from a blank page, and when a candidate is quietly in trouble the record shows it four months before anyone notices in conversation.
The screening pipeline
For candidates: six hundred abstracts against explicit inclusion and exclusion criteria written down first. Every decision recorded with the criterion that produced it, every borderline case flagged for the human. For systematic reviews this maps onto PRISMA and produces the flow numbers as a by-product.
The reproducibility spine
Data version, script version, environment, output. Every figure carries the command that made it. Set up in an afternoon in year one, and worth a month in year four.

Where it stops

It never generates a citation, and never summarises a paper you will cite.

Citations come from the reference manager and are resolved by DOI. Nothing that produces prose is permitted to produce a reference, ever, under any framing. Fabricated references are the single most common and most destructive failure mode of language models in research, and they are undetectable to a tired author at 23:00 because they look exactly right. Screening a thousand abstracts to find forty papers is machinery; reading those forty is the job — so screening produces titles, abstracts and links, never a synopsis presented as sufficient.

No manuscript under review touches an external service: peer review is confidential and a manuscript is unpublished work belonging to someone else. It never writes the argument, the hypothesis or the interpretation. It never ranks candidates in an appointment procedure — a Berufungsverfahren is legally reviewable, documented, and subject to the AGG; I have refused this twice. And no data is fabricated, no figure generated, and no cleaning happens silently: raw stays raw. Every engagement ends with a one-page document the client keeps, naming which tasks used a machine and which funder and publisher disclosure rules apply.

What proof looks like

Days from a call’s announcement to a submitted proposal: typically five weeks down to about ten working days. Proposals submitted per year. Supervision meetings with a written record: from occasional to all of them. Figures reproducible from a single command: all of them. And the only one anybody cares about — hours per week available for research, which for both professors went from close to zero to between eight and twelve.

When I say no

When someone wants a paper generator — asked twice, phrased as "a first draft of the introduction" and "just the related work section". The related work section is where you demonstrate you have read the field. Also when a department wants per-person publication metrics, which is surveillance in a bibliometric costume. And when a doctoral candidate needs a supervisor rather than a system, which is the saddest version.

What it costs you

Two weeks of extraction to build the core, once. Ninety seconds after each supervision meeting. And the discipline of reading the forty papers the screening found, which was always the job.

Kind words

Sixty per cent of every proposal already existed, in three earlier proposals, in slightly the wrong format, and I retyped it every March because finding it felt harder. Two weeks of extraction and I have not written those pages from nothing since. A DFG submission that took five weeks now takes about ten working days and nearly all of those days go to the work programme, which is the part that gets funded. He would not let anything generate a citation — references resolve by DOI or they do not appear — and given what I have seen happen to colleagues, that rule is the reason I let him near any of it.

Professor (W3), natural sciences

Year three, nine hundred unread PDFs, three chapters that no longer agreed with each other, and forty minutes with my supervisor every six weeks. The screening pipeline got six hundred abstracts down to the forty I actually had to read, and it recorded which criterion produced every decision. The reproducibility setup took an afternoon and I now know which script version made which figure, which I understand is a boring sentence and it saved me in my defence. He also told me, kindly, that what I really needed was a supervisor, and he was right.

Doctoral candidate, humanities

Get in touch

Tell me what your week looks like.

I'll tell you honestly which parts you can hand over — and which you can't.